Pattern Recognition by Neural-Network Processing in the Combined Sratial-Orientational Space

A dual stage neural-network architecture is proposed, for the combined local and global processing of visual dot-patterns, in the spatial-orientational space. Dipoles defining local orientation are first determined by a Hopfield-type network. The second stage, consisting of a layered neural network, is then trained by the back propagation algorithm with a new cost function to discriminate between dipole patterns. The network performance in classification tasks is similar to that of human vision.